The layer nobody owns

The handoff

Here is what work looks like right now on the teams I lead and at the companies I work with.

A developer sits down with a coding agent, sometimes two or three of them, each in its own harness. Together they write, test and refactor in an afternoon what used to take a sprint. Across the hall, or across the country, another developer is doing the same thing with their own agents, on the same codebase.

Then everything stops.

My developer’s agent can’t talk directly, and securely, to the other developer’s agent. It certainly can’t talk to the agents of a different team, a contractor or a vendor. So the work goes back through people: a pull request, a review, a message, an approval, a meeting to clear up what the message meant. The coding takes minutes. The handoff takes days.

The agents can write the code. They can’t talk to each other.

The stack we built for one person

It helps to look at what’s been built so far, because it’s easy to assume the plumbing is done.

There are four layers. The models themselves. The harness, which gives a model its arms and legs: tools, files, the ability to act. The evals and observability, which tell you whether it’s working and what it did. And the data underneath all of it.

Each of these layers is improving at a remarkable rate, and each was designed around the same picture: one person, using an agent, to do something against a system. Me and my agent, working in my codebase, my inbox, my company’s tools.

None of them answers the question my developers run into every day. What happens when two agents, acting for different people or different companies, need to work something out together?

Already the majority of traffic

The agents aren’t waiting for an answer.

Cloudflare’s CEO, Matthew Prince, had been forecasting when automated traffic would overtake human traffic on the internet. His company’s first estimate was the second half of 2027. Then the first half of 2027. In June he posted:

“Thought it would be end of 2027, then early 2027, but agentic traffic growing so fast that bots have now passed human traffic online for the first time in the Internet’s history.”

On Cloudflare’s measure, automated requests were about 57% of requests for web pages. Prince has since said he expects automated traffic to reach a thousand times human traffic within five years.

Two caveats, because they matter. Much of that traffic is still crawlers gathering training data, not agents doing work for someone, and some critics argued he picked a flattering metric. And almost none of it is agents talking to other agents. It’s agents working their way through a web built for people: reading pages, filling forms, clicking checkout buttons designed for a human thumb.

All of this is happening while most people barely use AI. As I wrote in the order of commitments, about half of American adults don’t use AI chatbots at all.

So here is my projection, and it is a projection. The volume of dealings between agents, the negotiating, scheduling, ordering, approving and settling, will be larger than anything people have handled before. Agents can be deployed at a scale, and work at a speed, that people can’t. That’s a long horizon. Given how far even Prince underestimated the last milestone, it may be shorter than anyone expects.

What exists, and what doesn’t

To be fair to the people building it, this layer isn’t empty. A common language exists and is maturing: open standards for how an agent reaches tools and data (MCP), and for how agents find and message each other (A2A). Both now sit under neutral foundations, with broad industry support. They’re real. They’re also young, and still rarely used across company lines in production.

A common language is not the same as a working relationship. What’s missing is the layer above it:

• Identity. Who, or what, is this agent?
• Authority. Whose agent is it, and what is it allowed to agree to on their behalf?
• Approvals and records. How does a commitment get approved, and recorded, in a way both companies will accept?
• Recourse. What happens when something goes wrong?

Agents posting messages to each other in a shared channel is not that layer. It’s a group chat with no rules, and everyone’s lawyers would agree.

Why nobody neutral owns it

The frontier labs and the large agent platforms are already enabling agents to work together, inside their own walls. If your agents and mine both live on the same platform, they can increasingly coordinate. That’s useful.

But the platforms have little reason to make it work across a rival’s platform. And even if they did, they wouldn’t be trusted to broker it. If my agent and yours disagree about what was agreed, neither of us wants the answer to come from a company with a stake in one side, or in keeping us both on its platform.

So the important question for the next few years isn’t only technical. It’s who earns the neutral seat.

Disputes are the tell

Here’s how you know this layer is real and not a whiteboard exercise: it will need a way to settle arguments.

When agents make commitments on our behalf, some of those commitments will go wrong. A price misread, a term one side thought was agreed, an approval that wasn’t actually given. Today that becomes an email thread, then a phone call, then possibly a lawyer. At agent speed and agent volume, that doesn’t scale. There has to be a neutral way to resolve disputes, ideally one the agents themselves can use, with people stepping in when it matters.

That is exactly the vision of the company whose advisory board I sit on. It’s also, I’d argue, the clearest sign of where this whole layer is heading. Commerce between strangers has always needed contracts, courts and arbitration. Commerce between agents will need its own version, and it will need it to be fast.

What it means for work

Go back to the developers. Now extend it: a company and its contractors building a system together, each side with its own agents. A vendor and a customer managing a process that crosses both their businesses. A procurement agent that needs a finance agent, at another company, to confirm something before anything moves.

In each case, the slowest part isn’t the work. It’s the handoff, the approval and the occasional dispute, and today all three run through people chasing each other over email. When those become things designed into how agents work together, the shape of the work changes. So do the questions that matter. Who is accountable when my agent agrees to something with yours? Where does a person still need to be in the loop, and where are they just slowing things down? What does it mean for your job when much of your counterpart is an agent?

A caution, before anyone rebuilds their org chart. All of this is long-horizon, and most of it still has to be proven in real use. Regulators are moving at their own pace, and so is adoption, both far slower than the technology. But once agent-to-agent work starts to take hold, I’d expect the pace of change to speed up even from where it is now, because agents improving the way agents work together compounds.

For investors

If you’re watching this space, three things are worth attention:

• Identity and delegated authority for agents. Proving who an agent is, and what it’s allowed to do for whom, is the foundation everything else depends on.
• Cross-company approvals and records. The paperwork of commerce, rebuilt for counterparties that are software.
• Neutral dispute resolution. The part everyone needs and no platform can credibly provide (admittedly a space where I have an interest).

The thread running through all three: in this layer, neutrality is the moat. That tilts the field toward independent players rather than platforms.

Back to the handoff

My developers will keep shipping code in an afternoon and waiting days for someone to approve it. For now.

The agents are ready to talk. Someone has to build the room they talk in, and it shouldn’t belong to either side of the table.